Unverified50% confidenceFactExact time
C51用51个等距原子离散化Q值分布支撑集，网络对每个动作输出51维softmax向量，训练目标为最小化预测分布与目标分布间的KL散度
1
Sources
50%
Confidence
Long-term
Relevance
8/20/2026
First Seen
Sources
Related Claims
UnverifiedQLoRA论文提出的NF4针对预训练权重服从正态分布的特性设计了最优量化格点63% similarUnverified后训练的最优解是参考分布的指数倾斜，即最优分布密度正比于参考模型密度乘以奖励的指数：π*(x) ∝ π_ref(x) · exp(r(x)/β)62% similarUnverified变分自编码器（VAE）和扩散模型的训练目标建立在变分推断和对数似然最大化的概率框架之上61% similarUnverifiedThe model training section of this project organizes full-pipeline resources covering pre-training, fine-tuning (SFT/RLHF/DPO), and efficient inference deployment.59% similarUnverifiedQuantization-Aware Training (QAT) lets the model retain a full-precision copy of weights, practicing rounding operations in each forward pass to learn weight distributions under extreme compression.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/778675API
curl https://kongchang.com/api/v1/knowledge/claims/778675MCP
get_claim(id=778675)